Researchers just proved that GitHub Copilot, Claude, and Gemini can be tricked into generating harmful code if you break the malicious request into steps and hide it inside normal development work.
The attack is simple in hindsight. A coder asks the assistant to read a file, then process some data, then output the result. All innocent steps. But the sequence of steps accomplishes something the model would refuse if asked directly. By the time the assistant realizes what is happening, it has already committed to the logic.
This matters because coding assistants are not just for writing better code anymore. They are touching databases, deploying infrastructure, accessing credentials. If the assistant gets compromised, the attacker has moved from stealing intellectual property to having hands inside production systems.
The labs are aware of this. They are working on it. But the tools are already in production and millions of developers are using them. Until the models get smarter about multi-step reasoning chains, this is the vulnerability surface we are working with.
——
Follow: @Ali Demi
Book your free AI clarity call, NOW!
https://buff.ly/TpWy277
——
Sources:
https://arxiv.org/pdf/2601.17548
https://www.techtimes.com/articles/320306/20260713/github-copilot-jailbreak-exploits-coding-workflow-bypass-all-safety-refusals.htm
https://www.scworld.com/brief/ai-coding-assistants-bypass-safety-filters-through-workflow-manipulation
Repost this. Thanks.

